Re: Thoughts on address selection

[Stig Venaas <stig at venaas.com>]

Scott Brim wrote:
> Stig Venaas allegedly wrote on 11/18/2009 1:44 PM:
> > Fred Baker wrote:
> > > On Nov 18, 2009, at 6:22 PM, Arifumi Matsumoto wrote:
> > >
> > > > I guess that is because if you force to try all the pairs, it perfectly
> > > > ignores the address selection manner defined in RFC 3484, and thus,
> > > > it gives us not little impact.
> > > If they space them closely and run them in parallel, I guess I don't
> > > see the impact. Imagine you have five addresses and your peer has five
> > > addresses, so there are 25 pairs. Imagine you are spacing the SYNs 10
> > > ms apart. Imagine that the only pair that works is the last one you try. 
> > I'm a bit worried about this. If e.g. the host is 100ms (RTT) away and
> > 10 combinations work, you may end up creating TCP state (and getting
> > syn-acks back) on the destination host for 10 connections, while you are
> > only going to use one.
>
> I think the general sense is to try pairs until you get a few acceptable
> ones, no more.

My point is that if say the spacing is 10ms and the RTT is 100ms, you
will needlessly be trying many pairs (up to 10) before you realize that
perhaps the first you tried actually works.

I would prefer the spacing to be >RTT so that you don't try a new
connection (with a new SYN) when the previous actually works.

Stig